DaGzang: a synthetic data generator for cross-domain recommendation services

Area of Science:

• Computer Science
• Artificial Intelligence
• Recommender Systems

Background:

• Cross-domain recommendation systems (CDRS) leverage inter-domain associations for improved user modeling and recommendations.
• A key challenge in CDRS is the absence of data for specific domains, hindering recommendation generation.
• Real-world identification of overlapping associations is difficult, limiting practical CDRS applications.

Purpose of the Study:

• To present DaGzang, a synthetic data generation platform specifically designed for cross-domain recommendation systems.
• To address the challenge of data scarcity in specific domains within CDRS.
• To facilitate the practical application of CDRS by overcoming difficulties in identifying real-world overlapping associations.

Main Methods:

• The DaGzang platform operates in a three-step loop: detecting overlap associations between real-world datasets, generating synthetic datasets based on these associations, and evaluating the synthetic data quality.
• Real-world datasets from Amazon's e-commerce platform were utilized for experiments.
• Synthetic datasets were integrated into the DakGalBi cross-domain recommender system for evaluation using collaborative filtering (CF) algorithms.

Main Results:

• The performance of CDRS using synthetic datasets was evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).
• User-based collaborative filtering (CF) demonstrated the highest performance among the evaluated methods.
• Optimal results were achieved with user-based CF utilizing 10 synthetic datasets generated by DaGzang, yielding MAE of 0.437 and RMSE of 0.465.

Conclusions:

• The DaGzang platform effectively generates synthetic datasets that enhance the performance of cross-domain recommendation systems.
• Synthetic data generation is a viable solution to address data sparsity issues in CDRS.
• User-based collaborative filtering shows strong efficacy when applied to synthetic datasets generated by the DaGzang platform.